This discussion is based on Yevgeni Berkovitch's Three big decisions. I hope people who are challenging the Google Analytics Competition for the first time get a lot of insight.
\n\ntransaction level vs user levelCurrently, there is a user level approach for oliver and a transaction level approach for others. \n- user_level_lightgbm LB 1.4480 olivier\n- [LB 1.4439]GACR Prediction: EDA + LGB Baseline by oskird
\n\nencoding method?Most people use one-hot-encoding and label-encoding. But there are a lot of other encoding techniques. How about using this to increase your LB?
\n\nclassification?The target variables of this competition are mostly 0, and the smaller number has a higher value. So classifying whether to buy or not to buy is a very important issue in raising your score.
\n\ntime-based-split vs k-fold ?I do not know in detail, but the conflict between the k-fold technique and thetime-based-split technique conflicts with the current validation division. If you know well with this part, please answer me.
Discrepancy between CV and LB scoresThere is a lot of difference between cv-score and lb-score, as you know who submitted it. There is a discussion on why this is the reason why and how to approach it.
\n\nother useful informationIt is the material that I arrange others documentation. because it was hard to keep searching. I hope it helps someone. Tell me if you can add anything.
", "messages": [ { "id": "396327", "postDate": "09/30/2018 13:22:45", "content": "This discussion is based on Yevgeni Berkovitch's Three big decisions. I hope people who are challenging the Google Analytics Competition for the first time get a lot of insight.
\n\ntransaction level vs user levelCurrently, there is a user level approach for oliver and a transaction level approach for others. \n- user_level_lightgbm LB 1.4480 olivier\n- [LB 1.4439]GACR Prediction: EDA + LGB Baseline by oskird
\n\nencoding method?Most people use one-hot-encoding and label-encoding. But there are a lot of other encoding techniques. How about using this to increase your LB?
\n\nclassification?The target variables of this competition are mostly 0, and the smaller number has a higher value. So classifying whether to buy or not to buy is a very important issue in raising your score.
\n\ntime-based-split vs k-fold ?I do not know in detail, but the conflict between the k-fold technique and thetime-based-split technique conflicts with the current validation division. If you know well with this part, please answer me.
Discrepancy between CV and LB scoresThere is a lot of difference between cv-score and lb-score, as you know who submitted it. There is a discussion on why this is the reason why and how to approach it.
\n\nother useful informationIt is the material that I arrange others documentation. because it was hard to keep searching. I hope it helps someone. Tell me if you can add anything.
", "rawMarkdown": "This discussion is based on Yevgeni Berkovitch's [Three big decisions](https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/66706). I hope people who are challenging the Google Analytics Competition for the first time get a lot of insight.\n\n## 1. `transaction level` vs `user level`\nCurrently, there is a user level approach for oliver and a transaction level approach for others. \n- [user_level_lightgbm LB 1.4480](https://www.kaggle.com/ogrellier/user-level-lightgbm-lb-1-4480) olivier\n- [[LB 1.4439]GACR Prediction: EDA + LGB Baseline](https://www.kaggle.com/c/ga-customer-revenue-prediction/kernels?sortBy=voteCount&group=everyone&pageSize=20&competitionId=10038&categoryIds=13102) by oskird\n\n## 2. what's your `encoding` method?\nMost people use one-hot-encoding and label-encoding. But there are a lot of other encoding techniques. How about using this to increase your LB? \n\n- [Various reference to solve high cardinality](https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/66581) by YouHan Lee\n- [Which encoding is good for time-validation? - 1.4417](https://www.kaggle.com/youhanlee/which-encoding-is-good-for-time-validation-1-4417) by YouHan Lee\n\n## 3. How about using `classification`? \nThe target variables of this competition are mostly 0, and the smaller number has a higher value. So classifying whether to buy or not to buy is a very important issue in raising your score.\n\n- [Could we use Classification ?\n](https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/66675) by olivier\n- [using-classification-for-predictions](https://www.kaggle.com/ogrellier/using-classification-for-predictions) by olivier\n\n## 4. `time-based-split` vs `k-fold` ?\nI do not know in detail, but the conflict between the `k-fold` technique and the` time-based-split` technique conflicts with the current validation division. If you know well with this part, please answer me.\n\n## 5. `Discrepancy between CV and LB scores`\nThere is a lot of difference between cv-score and lb-score, as you know who submitted it. There is a discussion on why this is the reason why and how to approach it.\n\n- [Discrepancy between CV and LB scores](https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/66811) by Craig Armstrong\n- [So where does the discrepancy between local scores and corrected LB come from?](https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/66739) by Adams\n\n## 6. `other useful information`\n- [Google Analytics columns description](https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/66046) by Alexander Firsov\n- [More than 10M single user revenue in test set](https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/66697) by AhmetErdem\n- [Will our models be useful?\n](https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/66440) by S D\n- [Date vs visitStartTime - explained](https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/66815#latest-395855) by Yannis Pappas\n- [[Finding] - Pageviews/Hits ratio <0.4 guarantees no revenue](https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/66774) by Adams\n- [Finding : totals_bounces and totals_transactionRevenue](https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/65989) by Maxwell\n\nIt is the material that I arrange others documentation. because it was hard to keep searching. I hope it helps someone. Tell me if you can add anything.", "votes": null }, { "id": "396805", "postDate": "10/01/2018 12:00:46", "content": "Oh, Thanks. Very good to see all the important discussions about this competition!
", "rawMarkdown": "Oh, Thanks. Very good to see all the important discussions about this competition!", "votes": null }, { "id": "397260", "postDate": "10/02/2018 07:51:44", "content": "Can you add a section with feature engineering ? cheers
", "rawMarkdown": "Can you add a section with feature engineering ? cheers", "votes": null }, { "id": "397401", "postDate": "10/02/2018 13:00:23", "content": "It's a good idea. However, features do not put people into the kernel or discussions because of their rankings. And need to verify that this is a good feature. yet, I'll try adding the proven features like pageviews / hits.. Thanks idea
haha. Thanks
", "rawMarkdown": "haha. Thanks", "votes": null } ], "comments": [ { "id": 396805, "author_name": "youhanlee", "author_url": "", "post_date": "10/01/2018 12:00:46", "content": "Oh, Thanks. Very good to see all the important discussions about this competition!
", "votes": null, "replies": [ { "id": 397402, "author_name": "chocozzz", "author_url": "", "post_date": "10/02/2018 13:00:35", "content": "haha. Thanks
", "votes": null, "replies": [] } ] }, { "id": 397260, "author_name": "xaviermaxime", "author_url": "", "post_date": "10/02/2018 07:51:44", "content": "Can you add a section with feature engineering ? cheers
", "votes": null, "replies": [ { "id": 397401, "author_name": "chocozzz", "author_url": "", "post_date": "10/02/2018 13:00:23", "content": "It's a good idea. However, features do not put people into the kernel or discussions because of their rankings. And need to verify that this is a good feature. yet, I'll try adding the proven features like pageviews / hits.. Thanks idea